Reconciliation is benchmarked near 11 hours per client each month, and almost none of it is matching. Where the time really lands, and which parts are checkable.
Bank reconciliation is the most measured task in a firm and one of the least examined. It appears on every job sheet, it consumes a predictable and substantial share of the month, and almost nobody breaks down what the hours are actually spent on.
One industry benchmark puts bank reconciliation at roughly 11.3 hours per client per month. At fifty clients that is around 565 hours — about three and a half full-time people, spent matching transactions. Whatever your own figure is, it is almost certainly measured in people rather than in hours, and it is almost certainly higher than the job sheets say, for reasons that become obvious once the work is broken up.
The instinctive response to that number is to look for a faster way to match transactions. That is the wrong target, because matching is not where the time goes.
The four things reconciliation is made of
Split the work honestly and it separates into four activities with very different characters.
Getting the statements. Requesting them, waiting, chasing, receiving them in the wrong format, asking again for the month that was missed, downloading from portals with expiring passwords. This is clerical, it is often not recorded against the job at all, and in most firms it is the largest single share. It is also the part nobody thinks of as reconciliation, which is precisely why it never gets measured.
Getting the numbers out of them. Reading PDFs into something structured. Where a bank feed exists this is close to free; where it does not, someone is reading a statement and typing, or running an extraction that has to be checked because a single misread digit invalidates everything downstream.
Matching. The activity the task is named after, and the one already largely automated in every serious ledger. Rules and feed matching clear the large majority without a person seeing them.
Investigating what did not match. The genuine work. A payment that appears twice, a transfer recorded as a cost, a refund whose sign was lost, a timing difference that is not actually a timing difference. Each requires judgement, and each is worth a qualified person’s attention.
The distribution matters more than the total. Firms that measure this typically find the first two categories — chasing and extraction — dominate, and that the fourth, the part that needs their qualification, is the smallest slice of an expensive month.
Chasing is a completeness problem in disguise
The reason statement collection consumes so much time is not that clients are slow, though they are. It is that nobody knows what is missing until late.
A folder of statements looks complete when it has a plausible number of files with consistent names. It can look exactly like that and be missing a period, and the omission surfaces when a reconciliation fails to tie, several days after the request that would have fixed it could have gone out.
Completeness is mechanically checkable, and it is checkable the moment the files arrive:
- Period continuity. Sort the statements by period and confirm the periods are contiguous with no gap and no overlap. This alone catches the missing month that a file count never will.
- Balance continuity. Each statement’s closing balance must equal the next statement’s opening balance. A break means either a missing statement or a misread figure, and the two are distinguishable by whether the gap is a whole period or a single value.
- Account coverage. Every account referenced anywhere in the ledger should have statements for every period in scope. Accounts get opened mid-year and quietly forgotten.
Running those three checks at intake converts a month of intermittent chasing into one accurate request, sent on day one, that names exactly what is missing. That is covered in more detail in is your bank statement set complete.
Extraction is cheap; unverified extraction is not
The second category has a subtler cost. Reading numbers out of a statement is easy to automate and easy to automate badly, and badly means a system that is usually right and cannot tell you when it is not.
A single transposed digit in an opening balance will propagate through an entire reconciliation and present as a difference that looks like a missing transaction. Someone will then spend an hour hunting a transaction that does not exist. The extraction saved twenty minutes and cost an hour, and the firm concludes that automation does not work here.
What makes extraction genuinely cheap is verification that does not depend on the extraction being right:
- Cross-engine agreement. Read each figure with two independent methods and compare. They fail differently, so where they agree, confidence is high; where they disagree, you have located the uncertainty precisely rather than discovering it downstream.
- Arithmetic the document asserts. Statements carry internal arithmetic — opening plus movements equals closing. Recompute it. If the recomputation does not hold, something was misread, and you know that before the figure reaches the ledger.
- Structural reading. Read by position on the page rather than by the order text appears in the file. Two-column statement layouts routinely interleave when read as a stream, which produces values attached to the wrong dates while every individual number is correct.
Each of these turns a silent error into a flagged exception. That is the whole difference between extraction that saves time and extraction that redistributes it.
Measuring your own split
The benchmark is only useful as an order of magnitude. What changes decisions is your own distribution, and it takes one month to establish rather than a project.
Pick five clients that differ — one with clean feeds, one with several PDF-only accounts, one habitually late with documents. For one cycle, record time against four buckets rather than one: requesting and chasing, extracting or keying, matching, and investigating breaks. Nothing else needs to change.
Two findings are common enough to predict. Chasing time is far larger than anyone expected, because it is spread across the month in short intervals that never get booked — three minutes writing an email, ten minutes a week later re-reading the thread to remember what was outstanding. And investigation, the part that genuinely needs a qualified person, is a small fraction of an expensive total.
That measurement is worth having before buying anything, including from us. It tells you which of the four categories is actually consuming your month, and therefore whether the constraint is collection, extraction or judgement — three different problems with three different answers. A firm whose time is going into chasing does not need better extraction, and will be disappointed by it.
It also gives you the only meaningful basis for evaluating a change afterwards. Without a before-figure, any improvement is a matter of impression, and impressions about time are unreliable in exactly the direction that makes automation look better than it was.
What is left is the work worth paying for
Once collection is one accurate request and extraction is verified at the point of reading, what remains on the desk is the exception list: the items that did not match and the reasons they did not.
That queue is small. It is also, uniformly, judgement — which is what the qualification is for. A senior person working through thirty flagged items with evidence attached is doing better work, and more of it per hour, than the same person working through four thousand rows to find those thirty.
None of this reduces the reconciliation. Every transaction still ties, every account still balances, and every decision is still made and signed by your firm. What changes is which hours are spent on the mechanical part — and at fifty clients, on a benchmark near eleven hours each, that difference is measured in people rather than in minutes.
This is the job we build as bank statement data extraction, and the same verification approach underpins month-end close automation.
Questions this raises
Our bank feeds import automatically. Does this still apply?
Feeds solve retrieval for the accounts that have them, which is a real saving. They do not solve completeness — a feed can silently miss a period or start mid-month — and they do not cover the accounts that still arrive as PDFs, which in most client portfolios is a meaningful minority. The continuity checks matter more with feeds, not less, because a gap in a feed is less visible than a missing file.
Is 11.3 hours per client realistic for our firm?
It is an industry benchmark, not a prediction about you. The reason to take it seriously is the order of magnitude rather than the decimal: at fifty clients, anything in that region is measured in full-time people. Firms that measure their own figure usually find it is higher than they assumed, because chasing time is rarely recorded against the job.
Which part of this can actually be automated?
Extraction and validation almost entirely, chasing partially, investigation not at all — and investigation is where the judgement lives, which is the right place for a qualified person to be spending time. The aim is not to remove the reconciliation. It is to have the hours land on the exceptions instead of on the clerical work in front of them.
What does the system do when it cannot read a figure?
Reports it as unresolved with the source shown, rather than guessing. An unresolved item is a valid output. A confidently wrong one is what makes people re-check everything, which costs more than doing the work by hand.